LLM Context Window Comparison

As of July 2026, the largest context window ever announced belongs to Magic's LTM-2-mini at 100,000,000 tokens (100M), but it is a research prototype that has never been publicly released.^[1] The biggest window on a model anyone can actually download is Meta's Llama 4 Scout, advertised at 10,000,000 tokens (10M); in practice every major host serves it at about 128,000 tokens, and Meta trained it on sequences no longer than 256K, so the 10M figure is an extrapolation rather than a validated recall length.^[2][3][4] Among production API models you can reliably call, the ceiling is roughly 1,000,000 tokens, a tier now shared by GPT-5.5, Claude Opus 4.8, Gemini 2.5 Pro, Grok 4.3, DeepSeek V4, MiniMax M3, Qwen3.7-Max and Llama 4 Maverick.^{[5][6][7][8][9][10][11][12]}

The single most important caveat is that advertised length is not usable length. Independent long-context benchmarks such as RULER and NoLiMa show that most models claiming 128K or more lose roughly half their effective recall well before that limit, and Anthropic now names the effect "context rot."^[13][14][15] For the underlying concepts (tokenization, attention scaling, RoPE, the lost-in-the-middle problem) see the main context window article; this page is the numbers-first comparison.

Which LLM has the biggest context window?

Ranked by advertised input tokens, and separated by what you can actually use:

1. Biggest ever announced: Magic LTM-2-mini, 100M tokens. Research-only, measured on Magic's own HashHop retrieval eval, not available to the public.^[1]
2. Biggest on a downloadable model: Llama 4 Scout, 10M advertised. Open-weight, but hosted deployments and Meta's own training cap the reliable length near 128K to 256K.^[2][3][4]
3. Biggest you can call at full length on a mainstream API: about 1M tokens, offered by GPT-5.5 (1,050,000), Gemini 2.5 Pro and Gemini 3.1 Pro (1,048,576), Claude Opus 4.8, Grok 4.3, DeepSeek V4, MiniMax M3, Qwen3.7-Max and Llama 4 Maverick (all 1,000,000).^{[5][6][7][8][9][10][11][12]}

Verdict: if your only metric is the advertised number, Llama 4 Scout (10M, open-weight) wins among usable models and Magic LTM-2-mini (100M) wins overall on paper. If you care about recall you can trust, treat any current model as reliable to roughly 128K to 200K and use retrieval beyond that; among 1M-token models, GPT-5.5 and Gemini 2.5 Pro currently hold their accuracy best at long range on the Fiction.liveBench comprehension test.^[16]

Full comparison table

Advertised context is the provider's stated maximum input in tokens. "Effective" is what benchmarks or hosts actually deliver. "Cost to fill" is the price of one input pass over the full advertised window at the input rate (top tier where pricing is length-tiered). Prices are USD per 1,000,000 tokens. Last verified: July 2026.

Scroll sideways for more →

Notes: (a) Scout cost-to-fill is theoretical: most hosts cap the served window near 128K, so you cannot actually submit 10M tokens. (b) GPT-5.5 (and GPT-5.4) double the input rate to $10.00 for the whole session once a prompt exceeds 272K tokens, so filling the window is billed at the higher tier.^[5] (c) Length-tiered: the higher rate applies to the entire prompt above the threshold (Gemini above 200K; MiniMax above 512K).^[7][9] (d) Cache-miss rate; a cache hit is far cheaper (DeepSeek V4-Pro about $0.0036/1M, Kimi K2.6 $0.16/1M input).^[10][12] Figures are drawn from each provider's official model card and pricing page (see References). Gemini 3.1 Pro Preview matches 2.5 Pro at 1,048,576 tokens but is priced higher at $2.00 / $4.00.^[8]

Advertised versus effective context: the recall gap

Every model above is trained to accept its advertised number of tokens, but accepting tokens is not the same as reasoning over them. Three lines of evidence make the gap concrete.

RULER, from NVIDIA, re-tests models at their claimed length with harder retrieval and tracking tasks. It finds that roughly half of the models advertising 128K or more fall below their short-context quality by 32K, and it assigns each an "effective context length" far below the label: GPT-4-1106 held 64K of its 128K claim, Command-R-plus held 32K of 128K, and Yi-34B held only 32K of a 200K claim.^[13] NoLiMa, from Adobe Research, removes the literal keyword overlap that lets a model cheat a search: under that harder setting, 11 of 13 tested long-context models dropped below 50% of their short-context score by 32K tokens, with GPT-4o falling from 99.3% to 69.7%.^[14] Fiction.liveBench, which tests story comprehension rather than keyword lookup, shows even strong 2025 models sliding from near-perfect at 8K to the 50% to 75% range by 120K, while the best recent models (the GPT-5 line and Gemini 2.5 Pro) retain high accuracy furthest into their windows.^[16]

Two structural reasons explain why the headline needle-in-a-haystack (NIAH) demos overstate real capability. First, NIAH plants a lexically distinctive sentence, so a model can win by string matching instead of comprehension; MiniMax reporting 100% NIAH accuracy at 4M tokens does not mean 4M tokens of usable reasoning.^[9][14] Second, attention is a finite budget: Anthropic's engineering guidance states that "as the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases," advising teams to treat context as "a precious, finite resource."^[15] For the mechanisms behind this (the lost-in-the-middle effect, attention scaling, position encodings such as RoPE and YaRN), see context window and LongBench.

What does it cost to fill a context window?

Filling a million-token prompt is not free, and the price gap between providers is enormous. At the input rates above, one full pass over the advertised window ranges from $0.20 on Claude Haiku 4.5 (200K) up to about $10.50 on GPT-5.5, whose input rate doubles to $10.00 per 1M once a prompt clears 272K tokens.^[5] Open-weight models are dramatically cheaper to fill: DeepSeek V4 costs about $0.44 for a 1M-token pass and Llama 4 Maverick about $0.50, versus $5.00 for Claude Opus 4.8 at the same length.^[7][10] Two pricing traps matter for long prompts. First, several providers charge more for long context: OpenAI applies a 2x input multiplier above 272K tokens, Google's Gemini Pro tier steps from $1.25 to $2.50 above 200K, and MiniMax steps from $0.30 to $0.60 above 512K, whereas Anthropic keeps the full 1M window at a single flat rate.^[5][7][9][6] Second, prompt caching changes the math: DeepSeek and Moonshot bill a cache hit at a fraction of the cache-miss rate, so re-using a large fixed context (a codebase, a document set) can be 5 to 100 times cheaper than paying to re-read it each call.^[10][12]

Open-weight versus proprietary long context

The very largest usable windows are now open-weight. Llama 4 Scout (10M advertised), DeepSeek V4 (MIT, 1M), MiniMax M3 (1M) and Kimi K2.6 (256K) can all be self-hosted, which removes per-token cost but shifts the burden to serving very long sequences efficiently, the reason hosts often cap Scout near 128K.^{[2][10][9][12]} Among proprietary models, the practical difference is less about the advertised number, which has converged near 1M, and more about retention and price: Gemini 2.5 Pro and GPT-5.5 currently show the best long-range recall, while Claude's flat 1M pricing and DeepSeek's cache economics make them the cheapest ways to actually use a large window.^[16][6][10] Historically notable open long-context milestones include MiniMax-Text-01 (4M advertised, January 2025), Qwen2.5-1M (the first open 1M model) and the now-retired Gemini 1.5 Pro, whose 2M window was the largest proprietary offering until Google consolidated every current Gemini model at 1,048,576 tokens.^[9][17][18]

References

1. Magic, "100M Token Context Windows," magic.dev/blog/100m-token-context-windows ↩
2. Meta AI, "The Llama 4 herd," ai.meta.com/blog/llama-4-multimodal-intelligence ↩
3. Hugging Face, model card "meta-llama/Llama-4-Scout-17B-16E," huggingface.co/meta-llama/Llama-4-Scout-17B-16E ↩
4. Groq, Llama 4 pricing and availability, groq.com/pricing and groq.com/blog/llama-4-now-live-on-groq ↩
5. OpenAI, GPT-5.5 model reference and API pricing, developers.openai.com/api/docs/models/gpt-5.5 and developers.openai.com/api/docs/pricing ↩
6. Anthropic, models overview and pricing, platform.claude.com/docs/en/about-claude/models/overview and platform.claude.com/docs/en/about-claude/pricing ↩
7. Google, Gemini 2.5 Pro model card and Gemini API pricing, ai.google.dev/gemini-api/docs/models/gemini-2.5-pro and ai.google.dev/gemini-api/docs/pricing ↩
8. Google, Gemini 3.1 Pro Preview model card, ai.google.dev/gemini-api/docs/models/gemini-3.1-pro-preview ↩
9. MiniMax, MiniMax-M3 model card and pay-as-you-go pricing, huggingface.co/MiniMaxAI/MiniMax-M3 and platform.minimax.io/docs/guides/pricing-paygo; MiniMax-01 (4M) arXiv:2501.08313 ↩
10. DeepSeek, "DeepSeek-V4 Preview" and API pricing, api-docs.deepseek.com/news/news260424 and api-docs.deepseek.com/quick_start/pricing; model card huggingface.co/deepseek-ai/DeepSeek-V4-Pro; arXiv:2606.19348 ↩
11. Alibaba Cloud Model Studio, Qwen3.7-Max model specification, alibabacloud.com/help/en/model-studio ↩
12. Moonshot AI, Kimi K2.6 model card and pricing, huggingface.co/moonshotai/Kimi-K2.6 and platform.kimi.ai/docs/pricing/chat-k26 ↩
13. Hsieh et al., "RULER: What's the Real Context Size of Your Long-Context Language Models?" arXiv:2404.06654; github.com/NVIDIA/RULER ↩
14. Modarressi et al., "NoLiMa: Long-Context Evaluation Beyond Literal Matching," arXiv:2502.05167 ↩
15. Anthropic, "Effective context engineering for AI agents," anthropic.com/engineering/effective-context-engineering-for-ai-agents ↩
16. Fiction.live, "Fiction.liveBench" long-context comprehension results, fiction.live and epoch.ai/benchmarks/fictionlivebench ↩
17. Alibaba, "Qwen2.5-1M: Deploy Your Own Qwen with Context Length up to 1M Tokens," qwenlm.github.io/blog/qwen2.5-1m; arXiv:2501.15383 ↩
18. Google, xAI Grok models documentation, docs.x.ai/docs/models; Google Gemini deprecations, ai.google.dev/gemini-api/docs/deprecations ↩